add: TSIR returns intenrnal variables as attr,

fix: numeric stable loss cals also for loss initialization

.gitignore

@@ -111,9 +111,8 @@ simulations/
 
 mlda_analysis/
 References/
-dataAnalysis/*
-!dataAnalysis/chess/
-dataAnalysis/chess/*.fen
+dataAnalysis/chess/*.Rdata
+dataAnalysis/Classification of EEG/
 
 *.csv
 *.csv.log

tensorPredictors/NAMESPACE

@@ -58,7 +58,6 @@ export(kpir.momentum)
 export(kpir.new)
 export(kronperm)
 export(mat)
-export(matProj)
 export(matpow)
 export(matrixImage)
 export(mcov)

tensorPredictors/R/TSIR.R

@@ -109,5 +109,5 @@ TSIR <- function(X, y, d, sample.axis = 1L,
 
     # reductions matrices `Omega_k^-1 Gamma_k` where there (reverse) kronecker
     # product spans the central tensor subspace (CTS) estimate
-    Map(solve, Omegas, Gammas)
+    structure(Map(solve, Omegas, Gammas), mcov = Omegas, Gammas = Gammas)
 }

tensorPredictors/R/gmlm_tensor_normal.R

@@ -61,8 +61,14 @@ gmlm_tensor_normal <- function(X, F, sample.axis = length(dim(X)),
     # Residuals
     R <- X - mlm(F, Map(`%*%`, Sigmas, betas))
 
+    # Numerically more stable version of `sum(log(mapply(det, Omegas)) / dimX)`
+    # which is itself equivalent to `log(det(Omega)) / prod(nrow(Omega))` where
+    # `Omega <- Reduce(kronecker, rev(Omegas))`.
+    det.Omega <- sum(mapply(function(Omega) {
+        sum(log(eigen(Omega, TRUE, TRUE)$values))
+    }, Omegas) / dimX)
     # Initial value of the log-likelihood (scaled and constants dropped)
-    loss <- mean(R * mlm(R, Omegas)) - sum(log(mapply(det, Omegas)) / dimX)
+    loss <- mean(R * mlm(R, Omegas)) - det.Omega
 
     # invoke the logger
     if (is.function(logger)) do.call(logger, list(